1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * Copyright (c) 1994 John S. Dyson 7 * All rights reserved. 8 * Copyright (c) 1994 David Greenman 9 * All rights reserved. 10 * 11 * 12 * This code is derived from software contributed to Berkeley by 13 * The Mach Operating System project at Carnegie-Mellon University. 14 * 15 * Redistribution and use in source and binary forms, with or without 16 * modification, are permitted provided that the following conditions 17 * are met: 18 * 1. Redistributions of source code must retain the above copyright 19 * notice, this list of conditions and the following disclaimer. 20 * 2. Redistributions in binary form must reproduce the above copyright 21 * notice, this list of conditions and the following disclaimer in the 22 * documentation and/or other materials provided with the distribution. 23 * 3. Neither the name of the University nor the names of its contributors 24 * may be used to endorse or promote products derived from this software 25 * without specific prior written permission. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 30 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 37 * SUCH DAMAGE. 38 * 39 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 40 * 41 * 42 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 43 * All rights reserved. 44 * 45 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 46 * 47 * Permission to use, copy, modify and distribute this software and 48 * its documentation is hereby granted, provided that both the copyright 49 * notice and this permission notice appear in all copies of the 50 * software, derivative works or modified versions, and any portions 51 * thereof, and that both notices appear in supporting documentation. 52 * 53 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 54 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 55 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 56 * 57 * Carnegie Mellon requests users of this software to return to 58 * 59 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 60 * School of Computer Science 61 * Carnegie Mellon University 62 * Pittsburgh PA 15213-3890 63 * 64 * any improvements or extensions that they make and grant Carnegie the 65 * rights to redistribute these changes. 66 * 67 * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $ 68 * $DragonFly: src/sys/vm/vm_fault.c,v 1.47 2008/07/01 02:02:56 dillon Exp $ 69 */ 70 71 /* 72 * Page fault handling module. 73 */ 74 75 #include <sys/param.h> 76 #include <sys/systm.h> 77 #include <sys/kernel.h> 78 #include <sys/proc.h> 79 #include <sys/vnode.h> 80 #include <sys/resourcevar.h> 81 #include <sys/vmmeter.h> 82 #include <sys/vkernel.h> 83 #include <sys/lock.h> 84 #include <sys/sysctl.h> 85 86 #include <cpu/lwbuf.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_param.h> 90 #include <vm/pmap.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_page.h> 94 #include <vm/vm_pageout.h> 95 #include <vm/vm_kern.h> 96 #include <vm/vm_pager.h> 97 #include <vm/vnode_pager.h> 98 #include <vm/vm_extern.h> 99 100 #include <sys/thread2.h> 101 #include <vm/vm_page2.h> 102 103 struct faultstate { 104 vm_page_t m; 105 vm_object_t object; 106 vm_pindex_t pindex; 107 vm_prot_t prot; 108 vm_page_t first_m; 109 vm_object_t first_object; 110 vm_prot_t first_prot; 111 vm_map_t map; 112 vm_map_entry_t entry; 113 int lookup_still_valid; 114 int hardfault; 115 int fault_flags; 116 int map_generation; 117 int shared; 118 int first_shared; 119 boolean_t wired; 120 struct vnode *vp; 121 }; 122 123 static int debug_fault = 0; 124 SYSCTL_INT(_vm, OID_AUTO, debug_fault, CTLFLAG_RW, &debug_fault, 0, ""); 125 static int debug_cluster = 0; 126 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, ""); 127 int vm_shared_fault = 1; 128 TUNABLE_INT("vm.shared_fault", &vm_shared_fault); 129 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW, &vm_shared_fault, 0, 130 "Allow shared token on vm_object"); 131 static long vm_shared_hit = 0; 132 SYSCTL_LONG(_vm, OID_AUTO, shared_hit, CTLFLAG_RW, &vm_shared_hit, 0, 133 "Successful shared faults"); 134 static long vm_shared_count = 0; 135 SYSCTL_LONG(_vm, OID_AUTO, shared_count, CTLFLAG_RW, &vm_shared_count, 0, 136 "Shared fault attempts"); 137 static long vm_shared_miss = 0; 138 SYSCTL_LONG(_vm, OID_AUTO, shared_miss, CTLFLAG_RW, &vm_shared_miss, 0, 139 "Unsuccessful shared faults"); 140 141 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int); 142 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, 143 vpte_t, int, int); 144 #if 0 145 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 146 #endif 147 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry); 148 static void vm_prefault(pmap_t pmap, vm_offset_t addra, 149 vm_map_entry_t entry, int prot, int fault_flags); 150 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 151 vm_map_entry_t entry, int prot, int fault_flags); 152 153 static __inline void 154 release_page(struct faultstate *fs) 155 { 156 vm_page_deactivate(fs->m); 157 vm_page_wakeup(fs->m); 158 fs->m = NULL; 159 } 160 161 /* 162 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse 163 * requires relocking and then checking the timestamp. 164 * 165 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do 166 * not have to update fs->map_generation here. 167 * 168 * NOTE: This function can fail due to a deadlock against the caller's 169 * holding of a vm_page BUSY. 170 */ 171 static __inline int 172 relock_map(struct faultstate *fs) 173 { 174 int error; 175 176 if (fs->lookup_still_valid == FALSE && fs->map) { 177 error = vm_map_lock_read_to(fs->map); 178 if (error == 0) 179 fs->lookup_still_valid = TRUE; 180 } else { 181 error = 0; 182 } 183 return error; 184 } 185 186 static __inline void 187 unlock_map(struct faultstate *fs) 188 { 189 if (fs->lookup_still_valid && fs->map) { 190 vm_map_lookup_done(fs->map, fs->entry, 0); 191 fs->lookup_still_valid = FALSE; 192 } 193 } 194 195 /* 196 * Clean up after a successful call to vm_fault_object() so another call 197 * to vm_fault_object() can be made. 198 */ 199 static void 200 _cleanup_successful_fault(struct faultstate *fs, int relock) 201 { 202 /* 203 * We allocated a junk page for a COW operation that did 204 * not occur, the page must be freed. 205 */ 206 if (fs->object != fs->first_object) { 207 KKASSERT(fs->first_shared == 0); 208 vm_page_free(fs->first_m); 209 vm_object_pip_wakeup(fs->object); 210 fs->first_m = NULL; 211 } 212 213 /* 214 * Reset fs->object. 215 */ 216 fs->object = fs->first_object; 217 if (relock && fs->lookup_still_valid == FALSE) { 218 if (fs->map) 219 vm_map_lock_read(fs->map); 220 fs->lookup_still_valid = TRUE; 221 } 222 } 223 224 static void 225 _unlock_things(struct faultstate *fs, int dealloc) 226 { 227 _cleanup_successful_fault(fs, 0); 228 if (dealloc) { 229 /*vm_object_deallocate(fs->first_object);*/ 230 /*fs->first_object = NULL; drop used later on */ 231 } 232 unlock_map(fs); 233 if (fs->vp != NULL) { 234 vput(fs->vp); 235 fs->vp = NULL; 236 } 237 } 238 239 #define unlock_things(fs) _unlock_things(fs, 0) 240 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 241 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1) 242 243 /* 244 * TRYPAGER 245 * 246 * Determine if the pager for the current object *might* contain the page. 247 * 248 * We only need to try the pager if this is not a default object (default 249 * objects are zero-fill and have no real pager), and if we are not taking 250 * a wiring fault or if the FS entry is wired. 251 */ 252 #define TRYPAGER(fs) \ 253 (fs->object->type != OBJT_DEFAULT && \ 254 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired)) 255 256 /* 257 * vm_fault: 258 * 259 * Handle a page fault occuring at the given address, requiring the given 260 * permissions, in the map specified. If successful, the page is inserted 261 * into the associated physical map. 262 * 263 * NOTE: The given address should be truncated to the proper page address. 264 * 265 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 266 * a standard error specifying why the fault is fatal is returned. 267 * 268 * The map in question must be referenced, and remains so. 269 * The caller may hold no locks. 270 * No other requirements. 271 */ 272 int 273 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 274 { 275 int result; 276 vm_pindex_t first_pindex; 277 struct faultstate fs; 278 struct lwp *lp; 279 int growstack; 280 int retry = 0; 281 282 vm_page_pcpu_cache(); 283 fs.hardfault = 0; 284 fs.fault_flags = fault_flags; 285 fs.vp = NULL; 286 fs.shared = vm_shared_fault; 287 fs.first_shared = vm_shared_fault; 288 growstack = 1; 289 if (vm_shared_fault) 290 ++vm_shared_count; 291 292 /* 293 * vm_map interactions 294 */ 295 if ((lp = curthread->td_lwp) != NULL) 296 lp->lwp_flags |= LWP_PAGING; 297 lwkt_gettoken(&map->token); 298 299 RetryFault: 300 /* 301 * Find the vm_map_entry representing the backing store and resolve 302 * the top level object and page index. This may have the side 303 * effect of executing a copy-on-write on the map entry and/or 304 * creating a shadow object, but will not COW any actual VM pages. 305 * 306 * On success fs.map is left read-locked and various other fields 307 * are initialized but not otherwise referenced or locked. 308 * 309 * NOTE! vm_map_lookup will try to upgrade the fault_type to 310 * VM_FAULT_WRITE if the map entry is a virtual page table and also 311 * writable, so we can set the 'A'accessed bit in the virtual page 312 * table entry. 313 */ 314 fs.map = map; 315 result = vm_map_lookup(&fs.map, vaddr, fault_type, 316 &fs.entry, &fs.first_object, 317 &first_pindex, &fs.first_prot, &fs.wired); 318 319 /* 320 * If the lookup failed or the map protections are incompatible, 321 * the fault generally fails. However, if the caller is trying 322 * to do a user wiring we have more work to do. 323 */ 324 if (result != KERN_SUCCESS) { 325 if (result != KERN_PROTECTION_FAILURE || 326 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 327 { 328 if (result == KERN_INVALID_ADDRESS && growstack && 329 map != &kernel_map && curproc != NULL) { 330 result = vm_map_growstack(curproc, vaddr); 331 if (result == KERN_SUCCESS) { 332 growstack = 0; 333 ++retry; 334 goto RetryFault; 335 } 336 result = KERN_FAILURE; 337 } 338 goto done; 339 } 340 341 /* 342 * If we are user-wiring a r/w segment, and it is COW, then 343 * we need to do the COW operation. Note that we don't 344 * currently COW RO sections now, because it is NOT desirable 345 * to COW .text. We simply keep .text from ever being COW'ed 346 * and take the heat that one cannot debug wired .text sections. 347 */ 348 result = vm_map_lookup(&fs.map, vaddr, 349 VM_PROT_READ|VM_PROT_WRITE| 350 VM_PROT_OVERRIDE_WRITE, 351 &fs.entry, &fs.first_object, 352 &first_pindex, &fs.first_prot, 353 &fs.wired); 354 if (result != KERN_SUCCESS) { 355 result = KERN_FAILURE; 356 goto done; 357 } 358 359 /* 360 * If we don't COW now, on a user wire, the user will never 361 * be able to write to the mapping. If we don't make this 362 * restriction, the bookkeeping would be nearly impossible. 363 * 364 * XXX We have a shared lock, this will have a MP race but 365 * I don't see how it can hurt anything. 366 */ 367 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 368 fs.entry->max_protection &= ~VM_PROT_WRITE; 369 } 370 371 /* 372 * fs.map is read-locked 373 * 374 * Misc checks. Save the map generation number to detect races. 375 */ 376 fs.map_generation = fs.map->timestamp; 377 fs.lookup_still_valid = TRUE; 378 fs.first_m = NULL; 379 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 380 381 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) { 382 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 383 panic("vm_fault: fault on nofault entry, addr: %p", 384 (void *)vaddr); 385 } 386 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) && 387 vaddr >= fs.entry->start && 388 vaddr < fs.entry->start + PAGE_SIZE) { 389 panic("vm_fault: fault on stack guard, addr: %p", 390 (void *)vaddr); 391 } 392 } 393 394 /* 395 * A system map entry may return a NULL object. No object means 396 * no pager means an unrecoverable kernel fault. 397 */ 398 if (fs.first_object == NULL) { 399 panic("vm_fault: unrecoverable fault at %p in entry %p", 400 (void *)vaddr, fs.entry); 401 } 402 403 /* 404 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 405 * is set. 406 */ 407 if ((curthread->td_flags & TDF_NOFAULT) && 408 (retry || 409 fs.first_object->type == OBJT_VNODE || 410 fs.first_object->backing_object)) { 411 result = KERN_FAILURE; 412 unlock_things(&fs); 413 goto done2; 414 } 415 416 /* 417 * If the entry is wired we cannot change the page protection. 418 */ 419 if (fs.wired) 420 fault_type = fs.first_prot; 421 422 /* 423 * We generally want to avoid unnecessary exclusive modes on backing 424 * and terminal objects because this can seriously interfere with 425 * heavily fork()'d processes (particularly /bin/sh scripts). 426 * 427 * However, we also want to avoid unnecessary retries due to needed 428 * shared->exclusive promotion for common faults. Exclusive mode is 429 * always needed if any page insertion, rename, or free occurs in an 430 * object (and also indirectly if any I/O is done). 431 * 432 * The main issue here is going to be fs.first_shared. If the 433 * first_object has a backing object which isn't shadowed and the 434 * process is single-threaded we might as well use an exclusive 435 * lock/chain right off the bat. 436 */ 437 if (fs.first_shared && fs.first_object->backing_object && 438 LIST_EMPTY(&fs.first_object->shadow_head) && 439 curthread->td_proc && curthread->td_proc->p_nthreads == 1) { 440 fs.first_shared = 0; 441 } 442 443 /* 444 * swap_pager_unswapped() needs an exclusive object 445 */ 446 if (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY)) { 447 fs.first_shared = 0; 448 } 449 450 /* 451 * Obtain a top-level object lock, shared or exclusive depending 452 * on fs.first_shared. If a shared lock winds up being insufficient 453 * we will retry with an exclusive lock. 454 * 455 * The vnode pager lock is always shared. 456 */ 457 if (fs.first_shared) 458 vm_object_hold_shared(fs.first_object); 459 else 460 vm_object_hold(fs.first_object); 461 if (fs.vp == NULL) 462 fs.vp = vnode_pager_lock(fs.first_object); 463 464 /* 465 * The page we want is at (first_object, first_pindex), but if the 466 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 467 * page table to figure out the actual pindex. 468 * 469 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 470 * ONLY 471 */ 472 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 473 result = vm_fault_vpagetable(&fs, &first_pindex, 474 fs.entry->aux.master_pde, 475 fault_type, 1); 476 if (result == KERN_TRY_AGAIN) { 477 vm_object_drop(fs.first_object); 478 ++retry; 479 goto RetryFault; 480 } 481 if (result != KERN_SUCCESS) 482 goto done; 483 } 484 485 /* 486 * Now we have the actual (object, pindex), fault in the page. If 487 * vm_fault_object() fails it will unlock and deallocate the FS 488 * data. If it succeeds everything remains locked and fs->object 489 * will have an additional PIP count if it is not equal to 490 * fs->first_object 491 * 492 * vm_fault_object will set fs->prot for the pmap operation. It is 493 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the 494 * page can be safely written. However, it will force a read-only 495 * mapping for a read fault if the memory is managed by a virtual 496 * page table. 497 * 498 * If the fault code uses the shared object lock shortcut 499 * we must not try to burst (we can't allocate VM pages). 500 */ 501 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 502 503 if (debug_fault > 0) { 504 --debug_fault; 505 kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x " 506 "fs.m=%p fs.prot=%02x fs.wired=%02x fs.entry=%p\n", 507 result, (intmax_t)vaddr, fault_type, fault_flags, 508 fs.m, fs.prot, fs.wired, fs.entry); 509 } 510 511 if (result == KERN_TRY_AGAIN) { 512 vm_object_drop(fs.first_object); 513 ++retry; 514 goto RetryFault; 515 } 516 if (result != KERN_SUCCESS) 517 goto done; 518 519 /* 520 * On success vm_fault_object() does not unlock or deallocate, and fs.m 521 * will contain a busied page. 522 * 523 * Enter the page into the pmap and do pmap-related adjustments. 524 */ 525 KKASSERT(fs.lookup_still_valid == TRUE); 526 vm_page_flag_set(fs.m, PG_REFERENCED); 527 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, fs.entry); 528 mycpu->gd_cnt.v_vm_faults++; 529 if (curthread->td_lwp) 530 ++curthread->td_lwp->lwp_ru.ru_minflt; 531 532 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */ 533 KKASSERT(fs.m->flags & PG_BUSY); 534 535 /* 536 * If the page is not wired down, then put it where the pageout daemon 537 * can find it. 538 */ 539 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 540 if (fs.wired) 541 vm_page_wire(fs.m); 542 else 543 vm_page_unwire(fs.m, 1); 544 } else { 545 vm_page_activate(fs.m); 546 } 547 vm_page_wakeup(fs.m); 548 549 /* 550 * Burst in a few more pages if possible. The fs.map should still 551 * be locked. To avoid interlocking against a vnode->getblk 552 * operation we had to be sure to unbusy our primary vm_page above 553 * first. 554 * 555 * A normal burst can continue down backing store, only execute 556 * if we are holding an exclusive lock, otherwise the exclusive 557 * locks the burst code gets might cause excessive SMP collisions. 558 * 559 * A quick burst can be utilized when there is no backing object 560 * (i.e. a shared file mmap). 561 */ 562 if ((fault_flags & VM_FAULT_BURST) && 563 (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 && 564 fs.wired == 0) { 565 if (fs.first_shared == 0 && fs.shared == 0) { 566 vm_prefault(fs.map->pmap, vaddr, 567 fs.entry, fs.prot, fault_flags); 568 } else { 569 vm_prefault_quick(fs.map->pmap, vaddr, 570 fs.entry, fs.prot, fault_flags); 571 } 572 } 573 574 /* 575 * Unlock everything, and return 576 */ 577 unlock_things(&fs); 578 579 if (curthread->td_lwp) { 580 if (fs.hardfault) { 581 curthread->td_lwp->lwp_ru.ru_majflt++; 582 } else { 583 curthread->td_lwp->lwp_ru.ru_minflt++; 584 } 585 } 586 587 /*vm_object_deallocate(fs.first_object);*/ 588 /*fs.m = NULL; */ 589 /*fs.first_object = NULL; must still drop later */ 590 591 result = KERN_SUCCESS; 592 done: 593 if (fs.first_object) 594 vm_object_drop(fs.first_object); 595 done2: 596 lwkt_reltoken(&map->token); 597 if (lp) 598 lp->lwp_flags &= ~LWP_PAGING; 599 if (vm_shared_fault && fs.shared == 0) 600 ++vm_shared_miss; 601 return (result); 602 } 603 604 /* 605 * Fault in the specified virtual address in the current process map, 606 * returning a held VM page or NULL. See vm_fault_page() for more 607 * information. 608 * 609 * No requirements. 610 */ 611 vm_page_t 612 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp) 613 { 614 struct lwp *lp = curthread->td_lwp; 615 vm_page_t m; 616 617 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 618 fault_type, VM_FAULT_NORMAL, errorp); 619 return(m); 620 } 621 622 /* 623 * Fault in the specified virtual address in the specified map, doing all 624 * necessary manipulation of the object store and all necessary I/O. Return 625 * a held VM page or NULL, and set *errorp. The related pmap is not 626 * updated. 627 * 628 * The returned page will be properly dirtied if VM_PROT_WRITE was specified, 629 * and marked PG_REFERENCED as well. 630 * 631 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an 632 * error will be returned. 633 * 634 * No requirements. 635 */ 636 vm_page_t 637 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 638 int fault_flags, int *errorp) 639 { 640 vm_pindex_t first_pindex; 641 struct faultstate fs; 642 int result; 643 int retry = 0; 644 vm_prot_t orig_fault_type = fault_type; 645 646 fs.hardfault = 0; 647 fs.fault_flags = fault_flags; 648 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 649 650 /* 651 * Dive the pmap (concurrency possible). If we find the 652 * appropriate page we can terminate early and quickly. 653 */ 654 fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type); 655 if (fs.m) { 656 *errorp = 0; 657 return(fs.m); 658 } 659 660 /* 661 * Otherwise take a concurrency hit and do a formal page 662 * fault. 663 */ 664 fs.shared = vm_shared_fault; 665 fs.first_shared = vm_shared_fault; 666 fs.vp = NULL; 667 lwkt_gettoken(&map->token); 668 669 /* 670 * swap_pager_unswapped() needs an exclusive object 671 */ 672 if (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY)) { 673 fs.first_shared = 0; 674 } 675 676 RetryFault: 677 /* 678 * Find the vm_map_entry representing the backing store and resolve 679 * the top level object and page index. This may have the side 680 * effect of executing a copy-on-write on the map entry and/or 681 * creating a shadow object, but will not COW any actual VM pages. 682 * 683 * On success fs.map is left read-locked and various other fields 684 * are initialized but not otherwise referenced or locked. 685 * 686 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 687 * if the map entry is a virtual page table and also writable, 688 * so we can set the 'A'accessed bit in the virtual page table entry. 689 */ 690 fs.map = map; 691 result = vm_map_lookup(&fs.map, vaddr, fault_type, 692 &fs.entry, &fs.first_object, 693 &first_pindex, &fs.first_prot, &fs.wired); 694 695 if (result != KERN_SUCCESS) { 696 *errorp = result; 697 fs.m = NULL; 698 goto done; 699 } 700 701 /* 702 * fs.map is read-locked 703 * 704 * Misc checks. Save the map generation number to detect races. 705 */ 706 fs.map_generation = fs.map->timestamp; 707 fs.lookup_still_valid = TRUE; 708 fs.first_m = NULL; 709 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 710 711 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 712 panic("vm_fault: fault on nofault entry, addr: %lx", 713 (u_long)vaddr); 714 } 715 716 /* 717 * A system map entry may return a NULL object. No object means 718 * no pager means an unrecoverable kernel fault. 719 */ 720 if (fs.first_object == NULL) { 721 panic("vm_fault: unrecoverable fault at %p in entry %p", 722 (void *)vaddr, fs.entry); 723 } 724 725 /* 726 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 727 * is set. 728 */ 729 if ((curthread->td_flags & TDF_NOFAULT) && 730 (retry || 731 fs.first_object->type == OBJT_VNODE || 732 fs.first_object->backing_object)) { 733 *errorp = KERN_FAILURE; 734 unlock_things(&fs); 735 goto done2; 736 } 737 738 /* 739 * If the entry is wired we cannot change the page protection. 740 */ 741 if (fs.wired) 742 fault_type = fs.first_prot; 743 744 /* 745 * Make a reference to this object to prevent its disposal while we 746 * are messing with it. Once we have the reference, the map is free 747 * to be diddled. Since objects reference their shadows (and copies), 748 * they will stay around as well. 749 * 750 * The reference should also prevent an unexpected collapse of the 751 * parent that might move pages from the current object into the 752 * parent unexpectedly, resulting in corruption. 753 * 754 * Bump the paging-in-progress count to prevent size changes (e.g. 755 * truncation operations) during I/O. This must be done after 756 * obtaining the vnode lock in order to avoid possible deadlocks. 757 */ 758 if (fs.first_shared) 759 vm_object_hold_shared(fs.first_object); 760 else 761 vm_object_hold(fs.first_object); 762 if (fs.vp == NULL) 763 fs.vp = vnode_pager_lock(fs.first_object); /* shared */ 764 765 /* 766 * The page we want is at (first_object, first_pindex), but if the 767 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 768 * page table to figure out the actual pindex. 769 * 770 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 771 * ONLY 772 */ 773 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 774 result = vm_fault_vpagetable(&fs, &first_pindex, 775 fs.entry->aux.master_pde, 776 fault_type, 1); 777 if (result == KERN_TRY_AGAIN) { 778 vm_object_drop(fs.first_object); 779 ++retry; 780 goto RetryFault; 781 } 782 if (result != KERN_SUCCESS) { 783 *errorp = result; 784 fs.m = NULL; 785 goto done; 786 } 787 } 788 789 /* 790 * Now we have the actual (object, pindex), fault in the page. If 791 * vm_fault_object() fails it will unlock and deallocate the FS 792 * data. If it succeeds everything remains locked and fs->object 793 * will have an additinal PIP count if it is not equal to 794 * fs->first_object 795 */ 796 fs.m = NULL; 797 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 798 799 if (result == KERN_TRY_AGAIN) { 800 vm_object_drop(fs.first_object); 801 ++retry; 802 goto RetryFault; 803 } 804 if (result != KERN_SUCCESS) { 805 *errorp = result; 806 fs.m = NULL; 807 goto done; 808 } 809 810 if ((orig_fault_type & VM_PROT_WRITE) && 811 (fs.prot & VM_PROT_WRITE) == 0) { 812 *errorp = KERN_PROTECTION_FAILURE; 813 unlock_and_deallocate(&fs); 814 fs.m = NULL; 815 goto done; 816 } 817 818 /* 819 * DO NOT UPDATE THE PMAP!!! This function may be called for 820 * a pmap unrelated to the current process pmap, in which case 821 * the current cpu core will not be listed in the pmap's pm_active 822 * mask. Thus invalidation interlocks will fail to work properly. 823 * 824 * (for example, 'ps' uses procfs to read program arguments from 825 * each process's stack). 826 * 827 * In addition to the above this function will be called to acquire 828 * a page that might already be faulted in, re-faulting it 829 * continuously is a waste of time. 830 * 831 * XXX could this have been the cause of our random seg-fault 832 * issues? procfs accesses user stacks. 833 */ 834 vm_page_flag_set(fs.m, PG_REFERENCED); 835 #if 0 836 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL); 837 mycpu->gd_cnt.v_vm_faults++; 838 if (curthread->td_lwp) 839 ++curthread->td_lwp->lwp_ru.ru_minflt; 840 #endif 841 842 /* 843 * On success vm_fault_object() does not unlock or deallocate, and fs.m 844 * will contain a busied page. So we must unlock here after having 845 * messed with the pmap. 846 */ 847 unlock_things(&fs); 848 849 /* 850 * Return a held page. We are not doing any pmap manipulation so do 851 * not set PG_MAPPED. However, adjust the page flags according to 852 * the fault type because the caller may not use a managed pmapping 853 * (so we don't want to lose the fact that the page will be dirtied 854 * if a write fault was specified). 855 */ 856 vm_page_hold(fs.m); 857 vm_page_activate(fs.m); 858 if (fault_type & VM_PROT_WRITE) 859 vm_page_dirty(fs.m); 860 861 if (curthread->td_lwp) { 862 if (fs.hardfault) { 863 curthread->td_lwp->lwp_ru.ru_majflt++; 864 } else { 865 curthread->td_lwp->lwp_ru.ru_minflt++; 866 } 867 } 868 869 /* 870 * Unlock everything, and return the held page. 871 */ 872 vm_page_wakeup(fs.m); 873 /*vm_object_deallocate(fs.first_object);*/ 874 /*fs.first_object = NULL; */ 875 *errorp = 0; 876 877 done: 878 if (fs.first_object) 879 vm_object_drop(fs.first_object); 880 done2: 881 lwkt_reltoken(&map->token); 882 return(fs.m); 883 } 884 885 /* 886 * Fault in the specified (object,offset), dirty the returned page as 887 * needed. If the requested fault_type cannot be done NULL and an 888 * error is returned. 889 * 890 * A held (but not busied) page is returned. 891 * 892 * The passed in object must be held as specified by the shared 893 * argument. 894 */ 895 vm_page_t 896 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset, 897 vm_prot_t fault_type, int fault_flags, 898 int *sharedp, int *errorp) 899 { 900 int result; 901 vm_pindex_t first_pindex; 902 struct faultstate fs; 903 struct vm_map_entry entry; 904 905 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 906 bzero(&entry, sizeof(entry)); 907 entry.object.vm_object = object; 908 entry.maptype = VM_MAPTYPE_NORMAL; 909 entry.protection = entry.max_protection = fault_type; 910 911 fs.hardfault = 0; 912 fs.fault_flags = fault_flags; 913 fs.map = NULL; 914 fs.shared = vm_shared_fault; 915 fs.first_shared = *sharedp; 916 fs.vp = NULL; 917 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 918 919 /* 920 * Might require swap block adjustments 921 */ 922 if (fs.first_shared && (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY))) { 923 fs.first_shared = 0; 924 vm_object_upgrade(object); 925 } 926 927 /* 928 * Retry loop as needed (typically for shared->exclusive transitions) 929 */ 930 RetryFault: 931 *sharedp = fs.first_shared; 932 first_pindex = OFF_TO_IDX(offset); 933 fs.first_object = object; 934 fs.entry = &entry; 935 fs.first_prot = fault_type; 936 fs.wired = 0; 937 /*fs.map_generation = 0; unused */ 938 939 /* 940 * Make a reference to this object to prevent its disposal while we 941 * are messing with it. Once we have the reference, the map is free 942 * to be diddled. Since objects reference their shadows (and copies), 943 * they will stay around as well. 944 * 945 * The reference should also prevent an unexpected collapse of the 946 * parent that might move pages from the current object into the 947 * parent unexpectedly, resulting in corruption. 948 * 949 * Bump the paging-in-progress count to prevent size changes (e.g. 950 * truncation operations) during I/O. This must be done after 951 * obtaining the vnode lock in order to avoid possible deadlocks. 952 */ 953 if (fs.vp == NULL) 954 fs.vp = vnode_pager_lock(fs.first_object); 955 956 fs.lookup_still_valid = TRUE; 957 fs.first_m = NULL; 958 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 959 960 #if 0 961 /* XXX future - ability to operate on VM object using vpagetable */ 962 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 963 result = vm_fault_vpagetable(&fs, &first_pindex, 964 fs.entry->aux.master_pde, 965 fault_type, 0); 966 if (result == KERN_TRY_AGAIN) { 967 if (fs.first_shared == 0 && *sharedp) 968 vm_object_upgrade(object); 969 goto RetryFault; 970 } 971 if (result != KERN_SUCCESS) { 972 *errorp = result; 973 return (NULL); 974 } 975 } 976 #endif 977 978 /* 979 * Now we have the actual (object, pindex), fault in the page. If 980 * vm_fault_object() fails it will unlock and deallocate the FS 981 * data. If it succeeds everything remains locked and fs->object 982 * will have an additinal PIP count if it is not equal to 983 * fs->first_object 984 * 985 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact. 986 * We may have to upgrade its lock to handle the requested fault. 987 */ 988 result = vm_fault_object(&fs, first_pindex, fault_type, 0); 989 990 if (result == KERN_TRY_AGAIN) { 991 if (fs.first_shared == 0 && *sharedp) 992 vm_object_upgrade(object); 993 goto RetryFault; 994 } 995 if (result != KERN_SUCCESS) { 996 *errorp = result; 997 return(NULL); 998 } 999 1000 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) { 1001 *errorp = KERN_PROTECTION_FAILURE; 1002 unlock_and_deallocate(&fs); 1003 return(NULL); 1004 } 1005 1006 /* 1007 * On success vm_fault_object() does not unlock or deallocate, so we 1008 * do it here. Note that the returned fs.m will be busied. 1009 */ 1010 unlock_things(&fs); 1011 1012 /* 1013 * Return a held page. We are not doing any pmap manipulation so do 1014 * not set PG_MAPPED. However, adjust the page flags according to 1015 * the fault type because the caller may not use a managed pmapping 1016 * (so we don't want to lose the fact that the page will be dirtied 1017 * if a write fault was specified). 1018 */ 1019 vm_page_hold(fs.m); 1020 vm_page_activate(fs.m); 1021 if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY)) 1022 vm_page_dirty(fs.m); 1023 if (fault_flags & VM_FAULT_UNSWAP) 1024 swap_pager_unswapped(fs.m); 1025 1026 /* 1027 * Indicate that the page was accessed. 1028 */ 1029 vm_page_flag_set(fs.m, PG_REFERENCED); 1030 1031 if (curthread->td_lwp) { 1032 if (fs.hardfault) { 1033 curthread->td_lwp->lwp_ru.ru_majflt++; 1034 } else { 1035 curthread->td_lwp->lwp_ru.ru_minflt++; 1036 } 1037 } 1038 1039 /* 1040 * Unlock everything, and return the held page. 1041 */ 1042 vm_page_wakeup(fs.m); 1043 /*vm_object_deallocate(fs.first_object);*/ 1044 /*fs.first_object = NULL; */ 1045 1046 *errorp = 0; 1047 return(fs.m); 1048 } 1049 1050 /* 1051 * Translate the virtual page number (first_pindex) that is relative 1052 * to the address space into a logical page number that is relative to the 1053 * backing object. Use the virtual page table pointed to by (vpte). 1054 * 1055 * This implements an N-level page table. Any level can terminate the 1056 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 1057 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 1058 */ 1059 static 1060 int 1061 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 1062 vpte_t vpte, int fault_type, int allow_nofault) 1063 { 1064 struct lwbuf *lwb; 1065 struct lwbuf lwb_cache; 1066 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */ 1067 int result = KERN_SUCCESS; 1068 vpte_t *ptep; 1069 1070 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1071 for (;;) { 1072 /* 1073 * We cannot proceed if the vpte is not valid, not readable 1074 * for a read fault, or not writable for a write fault. 1075 */ 1076 if ((vpte & VPTE_V) == 0) { 1077 unlock_and_deallocate(fs); 1078 return (KERN_FAILURE); 1079 } 1080 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) { 1081 unlock_and_deallocate(fs); 1082 return (KERN_FAILURE); 1083 } 1084 if ((vpte & VPTE_PS) || vshift == 0) 1085 break; 1086 KKASSERT(vshift >= VPTE_PAGE_BITS); 1087 1088 /* 1089 * Get the page table page. Nominally we only read the page 1090 * table, but since we are actively setting VPTE_M and VPTE_A, 1091 * tell vm_fault_object() that we are writing it. 1092 * 1093 * There is currently no real need to optimize this. 1094 */ 1095 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT, 1096 VM_PROT_READ|VM_PROT_WRITE, 1097 allow_nofault); 1098 if (result != KERN_SUCCESS) 1099 return (result); 1100 1101 /* 1102 * Process the returned fs.m and look up the page table 1103 * entry in the page table page. 1104 */ 1105 vshift -= VPTE_PAGE_BITS; 1106 lwb = lwbuf_alloc(fs->m, &lwb_cache); 1107 ptep = ((vpte_t *)lwbuf_kva(lwb) + 1108 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 1109 vpte = *ptep; 1110 1111 /* 1112 * Page table write-back. If the vpte is valid for the 1113 * requested operation, do a write-back to the page table. 1114 * 1115 * XXX VPTE_M is not set properly for page directory pages. 1116 * It doesn't get set in the page directory if the page table 1117 * is modified during a read access. 1118 */ 1119 vm_page_activate(fs->m); 1120 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) && 1121 (vpte & VPTE_RW)) { 1122 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) { 1123 atomic_set_long(ptep, VPTE_M | VPTE_A); 1124 vm_page_dirty(fs->m); 1125 } 1126 } 1127 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V)) { 1128 if ((vpte & VPTE_A) == 0) { 1129 atomic_set_long(ptep, VPTE_A); 1130 vm_page_dirty(fs->m); 1131 } 1132 } 1133 lwbuf_free(lwb); 1134 vm_page_flag_set(fs->m, PG_REFERENCED); 1135 vm_page_wakeup(fs->m); 1136 fs->m = NULL; 1137 cleanup_successful_fault(fs); 1138 } 1139 /* 1140 * Combine remaining address bits with the vpte. 1141 */ 1142 /* JG how many bits from each? */ 1143 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) + 1144 (*pindex & ((1L << vshift) - 1)); 1145 return (KERN_SUCCESS); 1146 } 1147 1148 1149 /* 1150 * This is the core of the vm_fault code. 1151 * 1152 * Do all operations required to fault-in (fs.first_object, pindex). Run 1153 * through the shadow chain as necessary and do required COW or virtual 1154 * copy operations. The caller has already fully resolved the vm_map_entry 1155 * and, if appropriate, has created a copy-on-write layer. All we need to 1156 * do is iterate the object chain. 1157 * 1158 * On failure (fs) is unlocked and deallocated and the caller may return or 1159 * retry depending on the failure code. On success (fs) is NOT unlocked or 1160 * deallocated, fs.m will contained a resolved, busied page, and fs.object 1161 * will have an additional PIP count if it is not equal to fs.first_object. 1162 * 1163 * If locks based on fs->first_shared or fs->shared are insufficient, 1164 * clear the appropriate field(s) and return RETRY. COWs require that 1165 * first_shared be 0, while page allocations (or frees) require that 1166 * shared be 0. Renames require that both be 0. 1167 * 1168 * fs->first_object must be held on call. 1169 */ 1170 static 1171 int 1172 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex, 1173 vm_prot_t fault_type, int allow_nofault) 1174 { 1175 vm_object_t next_object; 1176 vm_pindex_t pindex; 1177 int error; 1178 1179 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1180 fs->prot = fs->first_prot; 1181 fs->object = fs->first_object; 1182 pindex = first_pindex; 1183 1184 vm_object_chain_acquire(fs->first_object, fs->shared); 1185 vm_object_pip_add(fs->first_object, 1); 1186 1187 /* 1188 * If a read fault occurs we try to make the page writable if 1189 * possible. There are three cases where we cannot make the 1190 * page mapping writable: 1191 * 1192 * (1) The mapping is read-only or the VM object is read-only, 1193 * fs->prot above will simply not have VM_PROT_WRITE set. 1194 * 1195 * (2) If the mapping is a virtual page table we need to be able 1196 * to detect writes so we can set VPTE_M in the virtual page 1197 * table. 1198 * 1199 * (3) If the VM page is read-only or copy-on-write, upgrading would 1200 * just result in an unnecessary COW fault. 1201 * 1202 * VM_PROT_VPAGED is set if faulting via a virtual page table and 1203 * causes adjustments to the 'M'odify bit to also turn off write 1204 * access to force a re-fault. 1205 */ 1206 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1207 if ((fault_type & VM_PROT_WRITE) == 0) 1208 fs->prot &= ~VM_PROT_WRITE; 1209 } 1210 1211 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 1212 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 1213 if ((fault_type & VM_PROT_WRITE) == 0) 1214 fs->prot &= ~VM_PROT_WRITE; 1215 } 1216 1217 /* vm_object_hold(fs->object); implied b/c object == first_object */ 1218 1219 for (;;) { 1220 /* 1221 * The entire backing chain from first_object to object 1222 * inclusive is chainlocked. 1223 * 1224 * If the object is dead, we stop here 1225 */ 1226 if (fs->object->flags & OBJ_DEAD) { 1227 vm_object_pip_wakeup(fs->first_object); 1228 vm_object_chain_release_all(fs->first_object, 1229 fs->object); 1230 if (fs->object != fs->first_object) 1231 vm_object_drop(fs->object); 1232 unlock_and_deallocate(fs); 1233 return (KERN_PROTECTION_FAILURE); 1234 } 1235 1236 /* 1237 * See if the page is resident. Wait/Retry if the page is 1238 * busy (lots of stuff may have changed so we can't continue 1239 * in that case). 1240 * 1241 * We can theoretically allow the soft-busy case on a read 1242 * fault if the page is marked valid, but since such 1243 * pages are typically already pmap'd, putting that 1244 * special case in might be more effort then it is 1245 * worth. We cannot under any circumstances mess 1246 * around with a vm_page_t->busy page except, perhaps, 1247 * to pmap it. 1248 */ 1249 fs->m = vm_page_lookup_busy_try(fs->object, pindex, 1250 TRUE, &error); 1251 if (error) { 1252 vm_object_pip_wakeup(fs->first_object); 1253 vm_object_chain_release_all(fs->first_object, 1254 fs->object); 1255 if (fs->object != fs->first_object) 1256 vm_object_drop(fs->object); 1257 unlock_things(fs); 1258 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 1259 mycpu->gd_cnt.v_intrans++; 1260 /*vm_object_deallocate(fs->first_object);*/ 1261 /*fs->first_object = NULL;*/ 1262 fs->m = NULL; 1263 return (KERN_TRY_AGAIN); 1264 } 1265 if (fs->m) { 1266 /* 1267 * The page is busied for us. 1268 * 1269 * If reactivating a page from PQ_CACHE we may have 1270 * to rate-limit. 1271 */ 1272 int queue = fs->m->queue; 1273 vm_page_unqueue_nowakeup(fs->m); 1274 1275 if ((queue - fs->m->pc) == PQ_CACHE && 1276 vm_page_count_severe()) { 1277 vm_page_activate(fs->m); 1278 vm_page_wakeup(fs->m); 1279 fs->m = NULL; 1280 vm_object_pip_wakeup(fs->first_object); 1281 vm_object_chain_release_all(fs->first_object, 1282 fs->object); 1283 if (fs->object != fs->first_object) 1284 vm_object_drop(fs->object); 1285 unlock_and_deallocate(fs); 1286 if (allow_nofault == 0 || 1287 (curthread->td_flags & TDF_NOFAULT) == 0) { 1288 vm_wait_pfault(); 1289 } 1290 return (KERN_TRY_AGAIN); 1291 } 1292 1293 /* 1294 * If it still isn't completely valid (readable), 1295 * or if a read-ahead-mark is set on the VM page, 1296 * jump to readrest, else we found the page and 1297 * can return. 1298 * 1299 * We can release the spl once we have marked the 1300 * page busy. 1301 */ 1302 if (fs->m->object != &kernel_object) { 1303 if ((fs->m->valid & VM_PAGE_BITS_ALL) != 1304 VM_PAGE_BITS_ALL) { 1305 goto readrest; 1306 } 1307 if (fs->m->flags & PG_RAM) { 1308 if (debug_cluster) 1309 kprintf("R"); 1310 vm_page_flag_clear(fs->m, PG_RAM); 1311 goto readrest; 1312 } 1313 } 1314 break; /* break to PAGE HAS BEEN FOUND */ 1315 } 1316 1317 /* 1318 * Page is not resident, If this is the search termination 1319 * or the pager might contain the page, allocate a new page. 1320 */ 1321 if (TRYPAGER(fs) || fs->object == fs->first_object) { 1322 /* 1323 * Allocating, must be exclusive. 1324 */ 1325 if (fs->object == fs->first_object && 1326 fs->first_shared) { 1327 fs->first_shared = 0; 1328 vm_object_pip_wakeup(fs->first_object); 1329 vm_object_chain_release_all(fs->first_object, 1330 fs->object); 1331 if (fs->object != fs->first_object) 1332 vm_object_drop(fs->object); 1333 unlock_and_deallocate(fs); 1334 return (KERN_TRY_AGAIN); 1335 } 1336 if (fs->object != fs->first_object && 1337 fs->shared) { 1338 fs->first_shared = 0; 1339 fs->shared = 0; 1340 vm_object_pip_wakeup(fs->first_object); 1341 vm_object_chain_release_all(fs->first_object, 1342 fs->object); 1343 if (fs->object != fs->first_object) 1344 vm_object_drop(fs->object); 1345 unlock_and_deallocate(fs); 1346 return (KERN_TRY_AGAIN); 1347 } 1348 1349 /* 1350 * If the page is beyond the object size we fail 1351 */ 1352 if (pindex >= fs->object->size) { 1353 vm_object_pip_wakeup(fs->first_object); 1354 vm_object_chain_release_all(fs->first_object, 1355 fs->object); 1356 if (fs->object != fs->first_object) 1357 vm_object_drop(fs->object); 1358 unlock_and_deallocate(fs); 1359 return (KERN_PROTECTION_FAILURE); 1360 } 1361 1362 /* 1363 * Allocate a new page for this object/offset pair. 1364 * 1365 * It is possible for the allocation to race, so 1366 * handle the case. 1367 */ 1368 fs->m = NULL; 1369 if (!vm_page_count_severe()) { 1370 fs->m = vm_page_alloc(fs->object, pindex, 1371 ((fs->vp || fs->object->backing_object) ? 1372 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL : 1373 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1374 VM_ALLOC_USE_GD | VM_ALLOC_ZERO)); 1375 } 1376 if (fs->m == NULL) { 1377 vm_object_pip_wakeup(fs->first_object); 1378 vm_object_chain_release_all(fs->first_object, 1379 fs->object); 1380 if (fs->object != fs->first_object) 1381 vm_object_drop(fs->object); 1382 unlock_and_deallocate(fs); 1383 if (allow_nofault == 0 || 1384 (curthread->td_flags & TDF_NOFAULT) == 0) { 1385 vm_wait_pfault(); 1386 } 1387 return (KERN_TRY_AGAIN); 1388 } 1389 1390 /* 1391 * Fall through to readrest. We have a new page which 1392 * will have to be paged (since m->valid will be 0). 1393 */ 1394 } 1395 1396 readrest: 1397 /* 1398 * We have found an invalid or partially valid page, a 1399 * page with a read-ahead mark which might be partially or 1400 * fully valid (and maybe dirty too), or we have allocated 1401 * a new page. 1402 * 1403 * Attempt to fault-in the page if there is a chance that the 1404 * pager has it, and potentially fault in additional pages 1405 * at the same time. 1406 * 1407 * If TRYPAGER is true then fs.m will be non-NULL and busied 1408 * for us. 1409 */ 1410 if (TRYPAGER(fs)) { 1411 int rv; 1412 int seqaccess; 1413 u_char behavior = vm_map_entry_behavior(fs->entry); 1414 1415 if (behavior == MAP_ENTRY_BEHAV_RANDOM) 1416 seqaccess = 0; 1417 else 1418 seqaccess = -1; 1419 1420 /* 1421 * Doing I/O may synchronously insert additional 1422 * pages so we can't be shared at this point either. 1423 * 1424 * NOTE: We can't free fs->m here in the allocated 1425 * case (fs->object != fs->first_object) as 1426 * this would require an exclusively locked 1427 * VM object. 1428 */ 1429 if (fs->object == fs->first_object && 1430 fs->first_shared) { 1431 vm_page_deactivate(fs->m); 1432 vm_page_wakeup(fs->m); 1433 fs->m = NULL; 1434 fs->first_shared = 0; 1435 vm_object_pip_wakeup(fs->first_object); 1436 vm_object_chain_release_all(fs->first_object, 1437 fs->object); 1438 if (fs->object != fs->first_object) 1439 vm_object_drop(fs->object); 1440 unlock_and_deallocate(fs); 1441 return (KERN_TRY_AGAIN); 1442 } 1443 if (fs->object != fs->first_object && 1444 fs->shared) { 1445 vm_page_deactivate(fs->m); 1446 vm_page_wakeup(fs->m); 1447 fs->m = NULL; 1448 fs->first_shared = 0; 1449 fs->shared = 0; 1450 vm_object_pip_wakeup(fs->first_object); 1451 vm_object_chain_release_all(fs->first_object, 1452 fs->object); 1453 if (fs->object != fs->first_object) 1454 vm_object_drop(fs->object); 1455 unlock_and_deallocate(fs); 1456 return (KERN_TRY_AGAIN); 1457 } 1458 1459 /* 1460 * Avoid deadlocking against the map when doing I/O. 1461 * fs.object and the page is PG_BUSY'd. 1462 * 1463 * NOTE: Once unlocked, fs->entry can become stale 1464 * so this will NULL it out. 1465 * 1466 * NOTE: fs->entry is invalid until we relock the 1467 * map and verify that the timestamp has not 1468 * changed. 1469 */ 1470 unlock_map(fs); 1471 1472 /* 1473 * Acquire the page data. We still hold a ref on 1474 * fs.object and the page has been PG_BUSY's. 1475 * 1476 * The pager may replace the page (for example, in 1477 * order to enter a fictitious page into the 1478 * object). If it does so it is responsible for 1479 * cleaning up the passed page and properly setting 1480 * the new page PG_BUSY. 1481 * 1482 * If we got here through a PG_RAM read-ahead 1483 * mark the page may be partially dirty and thus 1484 * not freeable. Don't bother checking to see 1485 * if the pager has the page because we can't free 1486 * it anyway. We have to depend on the get_page 1487 * operation filling in any gaps whether there is 1488 * backing store or not. 1489 */ 1490 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess); 1491 1492 if (rv == VM_PAGER_OK) { 1493 /* 1494 * Relookup in case pager changed page. Pager 1495 * is responsible for disposition of old page 1496 * if moved. 1497 * 1498 * XXX other code segments do relookups too. 1499 * It's a bad abstraction that needs to be 1500 * fixed/removed. 1501 */ 1502 fs->m = vm_page_lookup(fs->object, pindex); 1503 if (fs->m == NULL) { 1504 vm_object_pip_wakeup(fs->first_object); 1505 vm_object_chain_release_all( 1506 fs->first_object, fs->object); 1507 if (fs->object != fs->first_object) 1508 vm_object_drop(fs->object); 1509 unlock_and_deallocate(fs); 1510 return (KERN_TRY_AGAIN); 1511 } 1512 ++fs->hardfault; 1513 break; /* break to PAGE HAS BEEN FOUND */ 1514 } 1515 1516 /* 1517 * Remove the bogus page (which does not exist at this 1518 * object/offset); before doing so, we must get back 1519 * our object lock to preserve our invariant. 1520 * 1521 * Also wake up any other process that may want to bring 1522 * in this page. 1523 * 1524 * If this is the top-level object, we must leave the 1525 * busy page to prevent another process from rushing 1526 * past us, and inserting the page in that object at 1527 * the same time that we are. 1528 */ 1529 if (rv == VM_PAGER_ERROR) { 1530 if (curproc) { 1531 kprintf("vm_fault: pager read error, " 1532 "pid %d (%s)\n", 1533 curproc->p_pid, 1534 curproc->p_comm); 1535 } else { 1536 kprintf("vm_fault: pager read error, " 1537 "thread %p (%s)\n", 1538 curthread, 1539 curproc->p_comm); 1540 } 1541 } 1542 1543 /* 1544 * Data outside the range of the pager or an I/O error 1545 * 1546 * The page may have been wired during the pagein, 1547 * e.g. by the buffer cache, and cannot simply be 1548 * freed. Call vnode_pager_freepage() to deal with it. 1549 * 1550 * Also note that we cannot free the page if we are 1551 * holding the related object shared. XXX not sure 1552 * what to do in that case. 1553 */ 1554 if (fs->object != fs->first_object) { 1555 vnode_pager_freepage(fs->m); 1556 fs->m = NULL; 1557 /* 1558 * XXX - we cannot just fall out at this 1559 * point, m has been freed and is invalid! 1560 */ 1561 } 1562 /* 1563 * XXX - the check for kernel_map is a kludge to work 1564 * around having the machine panic on a kernel space 1565 * fault w/ I/O error. 1566 */ 1567 if (((fs->map != &kernel_map) && 1568 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { 1569 if (fs->m) { 1570 if (fs->first_shared) { 1571 vm_page_deactivate(fs->m); 1572 vm_page_wakeup(fs->m); 1573 } else { 1574 vnode_pager_freepage(fs->m); 1575 } 1576 fs->m = NULL; 1577 } 1578 vm_object_pip_wakeup(fs->first_object); 1579 vm_object_chain_release_all(fs->first_object, 1580 fs->object); 1581 if (fs->object != fs->first_object) 1582 vm_object_drop(fs->object); 1583 unlock_and_deallocate(fs); 1584 if (rv == VM_PAGER_ERROR) 1585 return (KERN_FAILURE); 1586 else 1587 return (KERN_PROTECTION_FAILURE); 1588 /* NOT REACHED */ 1589 } 1590 } 1591 1592 /* 1593 * We get here if the object has a default pager (or unwiring) 1594 * or the pager doesn't have the page. 1595 * 1596 * fs->first_m will be used for the COW unless we find a 1597 * deeper page to be mapped read-only, in which case the 1598 * unlock*(fs) will free first_m. 1599 */ 1600 if (fs->object == fs->first_object) 1601 fs->first_m = fs->m; 1602 1603 /* 1604 * Move on to the next object. The chain lock should prevent 1605 * the backing_object from getting ripped out from under us. 1606 * 1607 * The object lock for the next object is governed by 1608 * fs->shared. 1609 */ 1610 if ((next_object = fs->object->backing_object) != NULL) { 1611 if (fs->shared) 1612 vm_object_hold_shared(next_object); 1613 else 1614 vm_object_hold(next_object); 1615 vm_object_chain_acquire(next_object, fs->shared); 1616 KKASSERT(next_object == fs->object->backing_object); 1617 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 1618 } 1619 1620 if (next_object == NULL) { 1621 /* 1622 * If there's no object left, fill the page in the top 1623 * object with zeros. 1624 */ 1625 if (fs->object != fs->first_object) { 1626 #if 0 1627 if (fs->first_object->backing_object != 1628 fs->object) { 1629 vm_object_hold(fs->first_object->backing_object); 1630 } 1631 #endif 1632 vm_object_chain_release_all( 1633 fs->first_object->backing_object, 1634 fs->object); 1635 #if 0 1636 if (fs->first_object->backing_object != 1637 fs->object) { 1638 vm_object_drop(fs->first_object->backing_object); 1639 } 1640 #endif 1641 vm_object_pip_wakeup(fs->object); 1642 vm_object_drop(fs->object); 1643 fs->object = fs->first_object; 1644 pindex = first_pindex; 1645 fs->m = fs->first_m; 1646 } 1647 fs->first_m = NULL; 1648 1649 /* 1650 * Zero the page if necessary and mark it valid. 1651 */ 1652 if ((fs->m->flags & PG_ZERO) == 0) { 1653 vm_page_zero_fill(fs->m); 1654 } else { 1655 #ifdef PMAP_DEBUG 1656 pmap_page_assertzero(VM_PAGE_TO_PHYS(fs->m)); 1657 #endif 1658 vm_page_flag_clear(fs->m, PG_ZERO); 1659 mycpu->gd_cnt.v_ozfod++; 1660 } 1661 mycpu->gd_cnt.v_zfod++; 1662 fs->m->valid = VM_PAGE_BITS_ALL; 1663 break; /* break to PAGE HAS BEEN FOUND */ 1664 } 1665 if (fs->object != fs->first_object) { 1666 vm_object_pip_wakeup(fs->object); 1667 vm_object_lock_swap(); 1668 vm_object_drop(fs->object); 1669 } 1670 KASSERT(fs->object != next_object, 1671 ("object loop %p", next_object)); 1672 fs->object = next_object; 1673 vm_object_pip_add(fs->object, 1); 1674 } 1675 1676 /* 1677 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1678 * is held.] 1679 * 1680 * object still held. 1681 * 1682 * local shared variable may be different from fs->shared. 1683 * 1684 * If the page is being written, but isn't already owned by the 1685 * top-level object, we have to copy it into a new page owned by the 1686 * top-level object. 1687 */ 1688 KASSERT((fs->m->flags & PG_BUSY) != 0, 1689 ("vm_fault: not busy after main loop")); 1690 1691 if (fs->object != fs->first_object) { 1692 /* 1693 * We only really need to copy if we want to write it. 1694 */ 1695 if (fault_type & VM_PROT_WRITE) { 1696 /* 1697 * This allows pages to be virtually copied from a 1698 * backing_object into the first_object, where the 1699 * backing object has no other refs to it, and cannot 1700 * gain any more refs. Instead of a bcopy, we just 1701 * move the page from the backing object to the 1702 * first object. Note that we must mark the page 1703 * dirty in the first object so that it will go out 1704 * to swap when needed. 1705 */ 1706 if ( 1707 /* 1708 * Must be holding exclusive locks 1709 */ 1710 fs->first_shared == 0 && 1711 fs->shared == 0 && 1712 /* 1713 * Map, if present, has not changed 1714 */ 1715 (fs->map == NULL || 1716 fs->map_generation == fs->map->timestamp) && 1717 /* 1718 * Only one shadow object 1719 */ 1720 (fs->object->shadow_count == 1) && 1721 /* 1722 * No COW refs, except us 1723 */ 1724 (fs->object->ref_count == 1) && 1725 /* 1726 * No one else can look this object up 1727 */ 1728 (fs->object->handle == NULL) && 1729 /* 1730 * No other ways to look the object up 1731 */ 1732 ((fs->object->type == OBJT_DEFAULT) || 1733 (fs->object->type == OBJT_SWAP)) && 1734 /* 1735 * We don't chase down the shadow chain 1736 */ 1737 (fs->object == fs->first_object->backing_object) && 1738 1739 /* 1740 * grab the lock if we need to 1741 */ 1742 (fs->lookup_still_valid || 1743 fs->map == NULL || 1744 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0) 1745 ) { 1746 /* 1747 * (first_m) and (m) are both busied. We have 1748 * move (m) into (first_m)'s object/pindex 1749 * in an atomic fashion, then free (first_m). 1750 * 1751 * first_object is held so second remove 1752 * followed by the rename should wind 1753 * up being atomic. vm_page_free() might 1754 * block so we don't do it until after the 1755 * rename. 1756 */ 1757 fs->lookup_still_valid = 1; 1758 vm_page_protect(fs->first_m, VM_PROT_NONE); 1759 vm_page_remove(fs->first_m); 1760 vm_page_rename(fs->m, fs->first_object, 1761 first_pindex); 1762 vm_page_free(fs->first_m); 1763 fs->first_m = fs->m; 1764 fs->m = NULL; 1765 mycpu->gd_cnt.v_cow_optim++; 1766 } else { 1767 /* 1768 * Oh, well, lets copy it. 1769 * 1770 * Why are we unmapping the original page 1771 * here? Well, in short, not all accessors 1772 * of user memory go through the pmap. The 1773 * procfs code doesn't have access user memory 1774 * via a local pmap, so vm_fault_page*() 1775 * can't call pmap_enter(). And the umtx*() 1776 * code may modify the COW'd page via a DMAP 1777 * or kernel mapping and not via the pmap, 1778 * leaving the original page still mapped 1779 * read-only into the pmap. 1780 * 1781 * So we have to remove the page from at 1782 * least the current pmap if it is in it. 1783 * Just remove it from all pmaps. 1784 */ 1785 KKASSERT(fs->first_shared == 0); 1786 vm_page_copy(fs->m, fs->first_m); 1787 vm_page_protect(fs->m, VM_PROT_NONE); 1788 vm_page_event(fs->m, VMEVENT_COW); 1789 } 1790 1791 /* 1792 * We no longer need the old page or object. 1793 */ 1794 if (fs->m) 1795 release_page(fs); 1796 1797 /* 1798 * We intend to revert to first_object, undo the 1799 * chain lock through to that. 1800 */ 1801 #if 0 1802 if (fs->first_object->backing_object != fs->object) 1803 vm_object_hold(fs->first_object->backing_object); 1804 #endif 1805 vm_object_chain_release_all( 1806 fs->first_object->backing_object, 1807 fs->object); 1808 #if 0 1809 if (fs->first_object->backing_object != fs->object) 1810 vm_object_drop(fs->first_object->backing_object); 1811 #endif 1812 1813 /* 1814 * fs->object != fs->first_object due to above 1815 * conditional 1816 */ 1817 vm_object_pip_wakeup(fs->object); 1818 vm_object_drop(fs->object); 1819 1820 /* 1821 * Only use the new page below... 1822 */ 1823 mycpu->gd_cnt.v_cow_faults++; 1824 fs->m = fs->first_m; 1825 fs->object = fs->first_object; 1826 pindex = first_pindex; 1827 } else { 1828 /* 1829 * If it wasn't a write fault avoid having to copy 1830 * the page by mapping it read-only. 1831 */ 1832 fs->prot &= ~VM_PROT_WRITE; 1833 } 1834 } 1835 1836 /* 1837 * Relock the map if necessary, then check the generation count. 1838 * relock_map() will update fs->timestamp to account for the 1839 * relocking if necessary. 1840 * 1841 * If the count has changed after relocking then all sorts of 1842 * crap may have happened and we have to retry. 1843 * 1844 * NOTE: The relock_map() can fail due to a deadlock against 1845 * the vm_page we are holding BUSY. 1846 */ 1847 if (fs->lookup_still_valid == FALSE && fs->map) { 1848 if (relock_map(fs) || 1849 fs->map->timestamp != fs->map_generation) { 1850 release_page(fs); 1851 vm_object_pip_wakeup(fs->first_object); 1852 vm_object_chain_release_all(fs->first_object, 1853 fs->object); 1854 if (fs->object != fs->first_object) 1855 vm_object_drop(fs->object); 1856 unlock_and_deallocate(fs); 1857 return (KERN_TRY_AGAIN); 1858 } 1859 } 1860 1861 /* 1862 * If the fault is a write, we know that this page is being 1863 * written NOW so dirty it explicitly to save on pmap_is_modified() 1864 * calls later. 1865 * 1866 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 1867 * if the page is already dirty to prevent data written with 1868 * the expectation of being synced from not being synced. 1869 * Likewise if this entry does not request NOSYNC then make 1870 * sure the page isn't marked NOSYNC. Applications sharing 1871 * data should use the same flags to avoid ping ponging. 1872 * 1873 * Also tell the backing pager, if any, that it should remove 1874 * any swap backing since the page is now dirty. 1875 */ 1876 vm_page_activate(fs->m); 1877 if (fs->prot & VM_PROT_WRITE) { 1878 vm_object_set_writeable_dirty(fs->m->object); 1879 vm_set_nosync(fs->m, fs->entry); 1880 if (fs->fault_flags & VM_FAULT_DIRTY) { 1881 vm_page_dirty(fs->m); 1882 swap_pager_unswapped(fs->m); 1883 } 1884 } 1885 1886 vm_object_pip_wakeup(fs->first_object); 1887 vm_object_chain_release_all(fs->first_object, fs->object); 1888 if (fs->object != fs->first_object) 1889 vm_object_drop(fs->object); 1890 1891 /* 1892 * Page had better still be busy. We are still locked up and 1893 * fs->object will have another PIP reference if it is not equal 1894 * to fs->first_object. 1895 */ 1896 KASSERT(fs->m->flags & PG_BUSY, 1897 ("vm_fault: page %p not busy!", fs->m)); 1898 1899 /* 1900 * Sanity check: page must be completely valid or it is not fit to 1901 * map into user space. vm_pager_get_pages() ensures this. 1902 */ 1903 if (fs->m->valid != VM_PAGE_BITS_ALL) { 1904 vm_page_zero_invalid(fs->m, TRUE); 1905 kprintf("Warning: page %p partially invalid on fault\n", fs->m); 1906 } 1907 vm_page_flag_clear(fs->m, PG_ZERO); 1908 1909 return (KERN_SUCCESS); 1910 } 1911 1912 /* 1913 * Hold each of the physical pages that are mapped by the specified range of 1914 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1915 * and allow the specified types of access, "prot". If all of the implied 1916 * pages are successfully held, then the number of held pages is returned 1917 * together with pointers to those pages in the array "ma". However, if any 1918 * of the pages cannot be held, -1 is returned. 1919 */ 1920 int 1921 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1922 vm_prot_t prot, vm_page_t *ma, int max_count) 1923 { 1924 vm_offset_t start, end; 1925 int i, npages, error; 1926 1927 start = trunc_page(addr); 1928 end = round_page(addr + len); 1929 1930 npages = howmany(end - start, PAGE_SIZE); 1931 1932 if (npages > max_count) 1933 return -1; 1934 1935 for (i = 0; i < npages; i++) { 1936 // XXX error handling 1937 ma[i] = vm_fault_page_quick(start + (i * PAGE_SIZE), 1938 prot, 1939 &error); 1940 } 1941 1942 return npages; 1943 } 1944 1945 /* 1946 * Wire down a range of virtual addresses in a map. The entry in question 1947 * should be marked in-transition and the map must be locked. We must 1948 * release the map temporarily while faulting-in the page to avoid a 1949 * deadlock. Note that the entry may be clipped while we are blocked but 1950 * will never be freed. 1951 * 1952 * No requirements. 1953 */ 1954 int 1955 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire) 1956 { 1957 boolean_t fictitious; 1958 vm_offset_t start; 1959 vm_offset_t end; 1960 vm_offset_t va; 1961 vm_paddr_t pa; 1962 vm_page_t m; 1963 pmap_t pmap; 1964 int rv; 1965 1966 lwkt_gettoken(&map->token); 1967 1968 pmap = vm_map_pmap(map); 1969 start = entry->start; 1970 end = entry->end; 1971 fictitious = entry->object.vm_object && 1972 ((entry->object.vm_object->type == OBJT_DEVICE) || 1973 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 1974 if (entry->eflags & MAP_ENTRY_KSTACK) 1975 start += PAGE_SIZE; 1976 map->timestamp++; 1977 vm_map_unlock(map); 1978 1979 /* 1980 * We simulate a fault to get the page and enter it in the physical 1981 * map. 1982 */ 1983 for (va = start; va < end; va += PAGE_SIZE) { 1984 if (user_wire) { 1985 rv = vm_fault(map, va, VM_PROT_READ, 1986 VM_FAULT_USER_WIRE); 1987 } else { 1988 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE, 1989 VM_FAULT_CHANGE_WIRING); 1990 } 1991 if (rv) { 1992 while (va > start) { 1993 va -= PAGE_SIZE; 1994 if ((pa = pmap_extract(pmap, va)) == 0) 1995 continue; 1996 pmap_change_wiring(pmap, va, FALSE, entry); 1997 if (!fictitious) { 1998 m = PHYS_TO_VM_PAGE(pa); 1999 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2000 vm_page_unwire(m, 1); 2001 vm_page_wakeup(m); 2002 } 2003 } 2004 goto done; 2005 } 2006 } 2007 rv = KERN_SUCCESS; 2008 done: 2009 vm_map_lock(map); 2010 lwkt_reltoken(&map->token); 2011 return (rv); 2012 } 2013 2014 /* 2015 * Unwire a range of virtual addresses in a map. The map should be 2016 * locked. 2017 */ 2018 void 2019 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 2020 { 2021 boolean_t fictitious; 2022 vm_offset_t start; 2023 vm_offset_t end; 2024 vm_offset_t va; 2025 vm_paddr_t pa; 2026 vm_page_t m; 2027 pmap_t pmap; 2028 2029 lwkt_gettoken(&map->token); 2030 2031 pmap = vm_map_pmap(map); 2032 start = entry->start; 2033 end = entry->end; 2034 fictitious = entry->object.vm_object && 2035 ((entry->object.vm_object->type == OBJT_DEVICE) || 2036 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 2037 if (entry->eflags & MAP_ENTRY_KSTACK) 2038 start += PAGE_SIZE; 2039 2040 /* 2041 * Since the pages are wired down, we must be able to get their 2042 * mappings from the physical map system. 2043 */ 2044 for (va = start; va < end; va += PAGE_SIZE) { 2045 pa = pmap_extract(pmap, va); 2046 if (pa != 0) { 2047 pmap_change_wiring(pmap, va, FALSE, entry); 2048 if (!fictitious) { 2049 m = PHYS_TO_VM_PAGE(pa); 2050 vm_page_busy_wait(m, FALSE, "vmwupg"); 2051 vm_page_unwire(m, 1); 2052 vm_page_wakeup(m); 2053 } 2054 } 2055 } 2056 lwkt_reltoken(&map->token); 2057 } 2058 2059 /* 2060 * Copy all of the pages from a wired-down map entry to another. 2061 * 2062 * The source and destination maps must be locked for write. 2063 * The source and destination maps token must be held 2064 * The source map entry must be wired down (or be a sharing map 2065 * entry corresponding to a main map entry that is wired down). 2066 * 2067 * No other requirements. 2068 * 2069 * XXX do segment optimization 2070 */ 2071 void 2072 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 2073 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 2074 { 2075 vm_object_t dst_object; 2076 vm_object_t src_object; 2077 vm_ooffset_t dst_offset; 2078 vm_ooffset_t src_offset; 2079 vm_prot_t prot; 2080 vm_offset_t vaddr; 2081 vm_page_t dst_m; 2082 vm_page_t src_m; 2083 2084 src_object = src_entry->object.vm_object; 2085 src_offset = src_entry->offset; 2086 2087 /* 2088 * Create the top-level object for the destination entry. (Doesn't 2089 * actually shadow anything - we copy the pages directly.) 2090 */ 2091 vm_map_entry_allocate_object(dst_entry); 2092 dst_object = dst_entry->object.vm_object; 2093 2094 prot = dst_entry->max_protection; 2095 2096 /* 2097 * Loop through all of the pages in the entry's range, copying each 2098 * one from the source object (it should be there) to the destination 2099 * object. 2100 */ 2101 vm_object_hold(src_object); 2102 vm_object_hold(dst_object); 2103 for (vaddr = dst_entry->start, dst_offset = 0; 2104 vaddr < dst_entry->end; 2105 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 2106 2107 /* 2108 * Allocate a page in the destination object 2109 */ 2110 do { 2111 dst_m = vm_page_alloc(dst_object, 2112 OFF_TO_IDX(dst_offset), 2113 VM_ALLOC_NORMAL); 2114 if (dst_m == NULL) { 2115 vm_wait(0); 2116 } 2117 } while (dst_m == NULL); 2118 2119 /* 2120 * Find the page in the source object, and copy it in. 2121 * (Because the source is wired down, the page will be in 2122 * memory.) 2123 */ 2124 src_m = vm_page_lookup(src_object, 2125 OFF_TO_IDX(dst_offset + src_offset)); 2126 if (src_m == NULL) 2127 panic("vm_fault_copy_wired: page missing"); 2128 2129 vm_page_copy(src_m, dst_m); 2130 vm_page_event(src_m, VMEVENT_COW); 2131 2132 /* 2133 * Enter it in the pmap... 2134 */ 2135 2136 vm_page_flag_clear(dst_m, PG_ZERO); 2137 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry); 2138 2139 /* 2140 * Mark it no longer busy, and put it on the active list. 2141 */ 2142 vm_page_activate(dst_m); 2143 vm_page_wakeup(dst_m); 2144 } 2145 vm_object_drop(dst_object); 2146 vm_object_drop(src_object); 2147 } 2148 2149 #if 0 2150 2151 /* 2152 * This routine checks around the requested page for other pages that 2153 * might be able to be faulted in. This routine brackets the viable 2154 * pages for the pages to be paged in. 2155 * 2156 * Inputs: 2157 * m, rbehind, rahead 2158 * 2159 * Outputs: 2160 * marray (array of vm_page_t), reqpage (index of requested page) 2161 * 2162 * Return value: 2163 * number of pages in marray 2164 */ 2165 static int 2166 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 2167 vm_page_t *marray, int *reqpage) 2168 { 2169 int i,j; 2170 vm_object_t object; 2171 vm_pindex_t pindex, startpindex, endpindex, tpindex; 2172 vm_page_t rtm; 2173 int cbehind, cahead; 2174 2175 object = m->object; 2176 pindex = m->pindex; 2177 2178 /* 2179 * we don't fault-ahead for device pager 2180 */ 2181 if ((object->type == OBJT_DEVICE) || 2182 (object->type == OBJT_MGTDEVICE)) { 2183 *reqpage = 0; 2184 marray[0] = m; 2185 return 1; 2186 } 2187 2188 /* 2189 * if the requested page is not available, then give up now 2190 */ 2191 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 2192 *reqpage = 0; /* not used by caller, fix compiler warn */ 2193 return 0; 2194 } 2195 2196 if ((cbehind == 0) && (cahead == 0)) { 2197 *reqpage = 0; 2198 marray[0] = m; 2199 return 1; 2200 } 2201 2202 if (rahead > cahead) { 2203 rahead = cahead; 2204 } 2205 2206 if (rbehind > cbehind) { 2207 rbehind = cbehind; 2208 } 2209 2210 /* 2211 * Do not do any readahead if we have insufficient free memory. 2212 * 2213 * XXX code was broken disabled before and has instability 2214 * with this conditonal fixed, so shortcut for now. 2215 */ 2216 if (burst_fault == 0 || vm_page_count_severe()) { 2217 marray[0] = m; 2218 *reqpage = 0; 2219 return 1; 2220 } 2221 2222 /* 2223 * scan backward for the read behind pages -- in memory 2224 * 2225 * Assume that if the page is not found an interrupt will not 2226 * create it. Theoretically interrupts can only remove (busy) 2227 * pages, not create new associations. 2228 */ 2229 if (pindex > 0) { 2230 if (rbehind > pindex) { 2231 rbehind = pindex; 2232 startpindex = 0; 2233 } else { 2234 startpindex = pindex - rbehind; 2235 } 2236 2237 vm_object_hold(object); 2238 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 2239 if (vm_page_lookup(object, tpindex - 1)) 2240 break; 2241 } 2242 2243 i = 0; 2244 while (tpindex < pindex) { 2245 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2246 VM_ALLOC_NULL_OK); 2247 if (rtm == NULL) { 2248 for (j = 0; j < i; j++) { 2249 vm_page_free(marray[j]); 2250 } 2251 vm_object_drop(object); 2252 marray[0] = m; 2253 *reqpage = 0; 2254 return 1; 2255 } 2256 marray[i] = rtm; 2257 ++i; 2258 ++tpindex; 2259 } 2260 vm_object_drop(object); 2261 } else { 2262 i = 0; 2263 } 2264 2265 /* 2266 * Assign requested page 2267 */ 2268 marray[i] = m; 2269 *reqpage = i; 2270 ++i; 2271 2272 /* 2273 * Scan forwards for read-ahead pages 2274 */ 2275 tpindex = pindex + 1; 2276 endpindex = tpindex + rahead; 2277 if (endpindex > object->size) 2278 endpindex = object->size; 2279 2280 vm_object_hold(object); 2281 while (tpindex < endpindex) { 2282 if (vm_page_lookup(object, tpindex)) 2283 break; 2284 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2285 VM_ALLOC_NULL_OK); 2286 if (rtm == NULL) 2287 break; 2288 marray[i] = rtm; 2289 ++i; 2290 ++tpindex; 2291 } 2292 vm_object_drop(object); 2293 2294 return (i); 2295 } 2296 2297 #endif 2298 2299 /* 2300 * vm_prefault() provides a quick way of clustering pagefaults into a 2301 * processes address space. It is a "cousin" of pmap_object_init_pt, 2302 * except it runs at page fault time instead of mmap time. 2303 * 2304 * vm.fast_fault Enables pre-faulting zero-fill pages 2305 * 2306 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to 2307 * prefault. Scan stops in either direction when 2308 * a page is found to already exist. 2309 * 2310 * This code used to be per-platform pmap_prefault(). It is now 2311 * machine-independent and enhanced to also pre-fault zero-fill pages 2312 * (see vm.fast_fault) as well as make them writable, which greatly 2313 * reduces the number of page faults programs incur. 2314 * 2315 * Application performance when pre-faulting zero-fill pages is heavily 2316 * dependent on the application. Very tiny applications like /bin/echo 2317 * lose a little performance while applications of any appreciable size 2318 * gain performance. Prefaulting multiple pages also reduces SMP 2319 * congestion and can improve SMP performance significantly. 2320 * 2321 * NOTE! prot may allow writing but this only applies to the top level 2322 * object. If we wind up mapping a page extracted from a backing 2323 * object we have to make sure it is read-only. 2324 * 2325 * NOTE! The caller has already handled any COW operations on the 2326 * vm_map_entry via the normal fault code. Do NOT call this 2327 * shortcut unless the normal fault code has run on this entry. 2328 * 2329 * The related map must be locked. 2330 * No other requirements. 2331 */ 2332 static int vm_prefault_pages = 8; 2333 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0, 2334 "Maximum number of pages to pre-fault"); 2335 static int vm_fast_fault = 1; 2336 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, 2337 "Burst fault zero-fill regions"); 2338 2339 /* 2340 * Set PG_NOSYNC if the map entry indicates so, but only if the page 2341 * is not already dirty by other means. This will prevent passive 2342 * filesystem syncing as well as 'sync' from writing out the page. 2343 */ 2344 static void 2345 vm_set_nosync(vm_page_t m, vm_map_entry_t entry) 2346 { 2347 if (entry->eflags & MAP_ENTRY_NOSYNC) { 2348 if (m->dirty == 0) 2349 vm_page_flag_set(m, PG_NOSYNC); 2350 } else { 2351 vm_page_flag_clear(m, PG_NOSYNC); 2352 } 2353 } 2354 2355 static void 2356 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot, 2357 int fault_flags) 2358 { 2359 struct lwp *lp; 2360 vm_page_t m; 2361 vm_offset_t addr; 2362 vm_pindex_t index; 2363 vm_pindex_t pindex; 2364 vm_object_t object; 2365 int pprot; 2366 int i; 2367 int noneg; 2368 int nopos; 2369 int maxpages; 2370 2371 /* 2372 * Get stable max count value, disabled if set to 0 2373 */ 2374 maxpages = vm_prefault_pages; 2375 cpu_ccfence(); 2376 if (maxpages <= 0) 2377 return; 2378 2379 /* 2380 * We do not currently prefault mappings that use virtual page 2381 * tables. We do not prefault foreign pmaps. 2382 */ 2383 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) 2384 return; 2385 lp = curthread->td_lwp; 2386 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2387 return; 2388 2389 /* 2390 * Limit pre-fault count to 1024 pages. 2391 */ 2392 if (maxpages > 1024) 2393 maxpages = 1024; 2394 2395 object = entry->object.vm_object; 2396 KKASSERT(object != NULL); 2397 KKASSERT(object == entry->object.vm_object); 2398 vm_object_hold(object); 2399 vm_object_chain_acquire(object, 0); 2400 2401 noneg = 0; 2402 nopos = 0; 2403 for (i = 0; i < maxpages; ++i) { 2404 vm_object_t lobject; 2405 vm_object_t nobject; 2406 int allocated = 0; 2407 int error; 2408 2409 /* 2410 * This can eat a lot of time on a heavily contended 2411 * machine so yield on the tick if needed. 2412 */ 2413 if ((i & 7) == 7) 2414 lwkt_yield(); 2415 2416 /* 2417 * Calculate the page to pre-fault, stopping the scan in 2418 * each direction separately if the limit is reached. 2419 */ 2420 if (i & 1) { 2421 if (noneg) 2422 continue; 2423 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2424 } else { 2425 if (nopos) 2426 continue; 2427 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2428 } 2429 if (addr < entry->start) { 2430 noneg = 1; 2431 if (noneg && nopos) 2432 break; 2433 continue; 2434 } 2435 if (addr >= entry->end) { 2436 nopos = 1; 2437 if (noneg && nopos) 2438 break; 2439 continue; 2440 } 2441 2442 /* 2443 * Skip pages already mapped, and stop scanning in that 2444 * direction. When the scan terminates in both directions 2445 * we are done. 2446 */ 2447 if (pmap_prefault_ok(pmap, addr) == 0) { 2448 if (i & 1) 2449 noneg = 1; 2450 else 2451 nopos = 1; 2452 if (noneg && nopos) 2453 break; 2454 continue; 2455 } 2456 2457 /* 2458 * Follow the VM object chain to obtain the page to be mapped 2459 * into the pmap. 2460 * 2461 * If we reach the terminal object without finding a page 2462 * and we determine it would be advantageous, then allocate 2463 * a zero-fill page for the base object. The base object 2464 * is guaranteed to be OBJT_DEFAULT for this case. 2465 * 2466 * In order to not have to check the pager via *haspage*() 2467 * we stop if any non-default object is encountered. e.g. 2468 * a vnode or swap object would stop the loop. 2469 */ 2470 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2471 lobject = object; 2472 pindex = index; 2473 pprot = prot; 2474 2475 KKASSERT(lobject == entry->object.vm_object); 2476 /*vm_object_hold(lobject); implied */ 2477 2478 while ((m = vm_page_lookup_busy_try(lobject, pindex, 2479 TRUE, &error)) == NULL) { 2480 if (lobject->type != OBJT_DEFAULT) 2481 break; 2482 if (lobject->backing_object == NULL) { 2483 if (vm_fast_fault == 0) 2484 break; 2485 if ((prot & VM_PROT_WRITE) == 0 || 2486 vm_page_count_min(0)) { 2487 break; 2488 } 2489 2490 /* 2491 * NOTE: Allocated from base object 2492 */ 2493 m = vm_page_alloc(object, index, 2494 VM_ALLOC_NORMAL | 2495 VM_ALLOC_ZERO | 2496 VM_ALLOC_USE_GD | 2497 VM_ALLOC_NULL_OK); 2498 if (m == NULL) 2499 break; 2500 allocated = 1; 2501 pprot = prot; 2502 /* lobject = object .. not needed */ 2503 break; 2504 } 2505 if (lobject->backing_object_offset & PAGE_MASK) 2506 break; 2507 nobject = lobject->backing_object; 2508 vm_object_hold(nobject); 2509 KKASSERT(nobject == lobject->backing_object); 2510 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 2511 if (lobject != object) { 2512 vm_object_lock_swap(); 2513 vm_object_drop(lobject); 2514 } 2515 lobject = nobject; 2516 pprot &= ~VM_PROT_WRITE; 2517 vm_object_chain_acquire(lobject, 0); 2518 } 2519 2520 /* 2521 * NOTE: A non-NULL (m) will be associated with lobject if 2522 * it was found there, otherwise it is probably a 2523 * zero-fill page associated with the base object. 2524 * 2525 * Give-up if no page is available. 2526 */ 2527 if (m == NULL) { 2528 if (lobject != object) { 2529 #if 0 2530 if (object->backing_object != lobject) 2531 vm_object_hold(object->backing_object); 2532 #endif 2533 vm_object_chain_release_all( 2534 object->backing_object, lobject); 2535 #if 0 2536 if (object->backing_object != lobject) 2537 vm_object_drop(object->backing_object); 2538 #endif 2539 vm_object_drop(lobject); 2540 } 2541 break; 2542 } 2543 2544 /* 2545 * The object must be marked dirty if we are mapping a 2546 * writable page. m->object is either lobject or object, 2547 * both of which are still held. Do this before we 2548 * potentially drop the object. 2549 */ 2550 if (pprot & VM_PROT_WRITE) 2551 vm_object_set_writeable_dirty(m->object); 2552 2553 /* 2554 * Do not conditionalize on PG_RAM. If pages are present in 2555 * the VM system we assume optimal caching. If caching is 2556 * not optimal the I/O gravy train will be restarted when we 2557 * hit an unavailable page. We do not want to try to restart 2558 * the gravy train now because we really don't know how much 2559 * of the object has been cached. The cost for restarting 2560 * the gravy train should be low (since accesses will likely 2561 * be I/O bound anyway). 2562 */ 2563 if (lobject != object) { 2564 #if 0 2565 if (object->backing_object != lobject) 2566 vm_object_hold(object->backing_object); 2567 #endif 2568 vm_object_chain_release_all(object->backing_object, 2569 lobject); 2570 #if 0 2571 if (object->backing_object != lobject) 2572 vm_object_drop(object->backing_object); 2573 #endif 2574 vm_object_drop(lobject); 2575 } 2576 2577 /* 2578 * Enter the page into the pmap if appropriate. If we had 2579 * allocated the page we have to place it on a queue. If not 2580 * we just have to make sure it isn't on the cache queue 2581 * (pages on the cache queue are not allowed to be mapped). 2582 */ 2583 if (allocated) { 2584 /* 2585 * Page must be zerod. 2586 */ 2587 if ((m->flags & PG_ZERO) == 0) { 2588 vm_page_zero_fill(m); 2589 } else { 2590 #ifdef PMAP_DEBUG 2591 pmap_page_assertzero( 2592 VM_PAGE_TO_PHYS(m)); 2593 #endif 2594 vm_page_flag_clear(m, PG_ZERO); 2595 mycpu->gd_cnt.v_ozfod++; 2596 } 2597 mycpu->gd_cnt.v_zfod++; 2598 m->valid = VM_PAGE_BITS_ALL; 2599 2600 /* 2601 * Handle dirty page case 2602 */ 2603 if (pprot & VM_PROT_WRITE) 2604 vm_set_nosync(m, entry); 2605 pmap_enter(pmap, addr, m, pprot, 0, entry); 2606 mycpu->gd_cnt.v_vm_faults++; 2607 if (curthread->td_lwp) 2608 ++curthread->td_lwp->lwp_ru.ru_minflt; 2609 vm_page_deactivate(m); 2610 if (pprot & VM_PROT_WRITE) { 2611 /*vm_object_set_writeable_dirty(m->object);*/ 2612 vm_set_nosync(m, entry); 2613 if (fault_flags & VM_FAULT_DIRTY) { 2614 vm_page_dirty(m); 2615 /*XXX*/ 2616 swap_pager_unswapped(m); 2617 } 2618 } 2619 vm_page_wakeup(m); 2620 } else if (error) { 2621 /* couldn't busy page, no wakeup */ 2622 } else if ( 2623 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2624 (m->flags & PG_FICTITIOUS) == 0) { 2625 /* 2626 * A fully valid page not undergoing soft I/O can 2627 * be immediately entered into the pmap. 2628 */ 2629 if ((m->queue - m->pc) == PQ_CACHE) 2630 vm_page_deactivate(m); 2631 if (pprot & VM_PROT_WRITE) { 2632 /*vm_object_set_writeable_dirty(m->object);*/ 2633 vm_set_nosync(m, entry); 2634 if (fault_flags & VM_FAULT_DIRTY) { 2635 vm_page_dirty(m); 2636 /*XXX*/ 2637 swap_pager_unswapped(m); 2638 } 2639 } 2640 if (pprot & VM_PROT_WRITE) 2641 vm_set_nosync(m, entry); 2642 pmap_enter(pmap, addr, m, pprot, 0, entry); 2643 mycpu->gd_cnt.v_vm_faults++; 2644 if (curthread->td_lwp) 2645 ++curthread->td_lwp->lwp_ru.ru_minflt; 2646 vm_page_wakeup(m); 2647 } else { 2648 vm_page_wakeup(m); 2649 } 2650 } 2651 vm_object_chain_release(object); 2652 vm_object_drop(object); 2653 } 2654 2655 /* 2656 * Object can be held shared 2657 */ 2658 static void 2659 vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 2660 vm_map_entry_t entry, int prot, int fault_flags) 2661 { 2662 struct lwp *lp; 2663 vm_page_t m; 2664 vm_offset_t addr; 2665 vm_pindex_t pindex; 2666 vm_object_t object; 2667 int i; 2668 int noneg; 2669 int nopos; 2670 int maxpages; 2671 2672 /* 2673 * Get stable max count value, disabled if set to 0 2674 */ 2675 maxpages = vm_prefault_pages; 2676 cpu_ccfence(); 2677 if (maxpages <= 0) 2678 return; 2679 2680 /* 2681 * We do not currently prefault mappings that use virtual page 2682 * tables. We do not prefault foreign pmaps. 2683 */ 2684 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) 2685 return; 2686 lp = curthread->td_lwp; 2687 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2688 return; 2689 object = entry->object.vm_object; 2690 if (object->backing_object != NULL) 2691 return; 2692 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2693 2694 /* 2695 * Limit pre-fault count to 1024 pages. 2696 */ 2697 if (maxpages > 1024) 2698 maxpages = 1024; 2699 2700 noneg = 0; 2701 nopos = 0; 2702 for (i = 0; i < maxpages; ++i) { 2703 int error; 2704 2705 /* 2706 * Calculate the page to pre-fault, stopping the scan in 2707 * each direction separately if the limit is reached. 2708 */ 2709 if (i & 1) { 2710 if (noneg) 2711 continue; 2712 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2713 } else { 2714 if (nopos) 2715 continue; 2716 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2717 } 2718 if (addr < entry->start) { 2719 noneg = 1; 2720 if (noneg && nopos) 2721 break; 2722 continue; 2723 } 2724 if (addr >= entry->end) { 2725 nopos = 1; 2726 if (noneg && nopos) 2727 break; 2728 continue; 2729 } 2730 2731 /* 2732 * Skip pages already mapped, and stop scanning in that 2733 * direction. When the scan terminates in both directions 2734 * we are done. 2735 */ 2736 if (pmap_prefault_ok(pmap, addr) == 0) { 2737 if (i & 1) 2738 noneg = 1; 2739 else 2740 nopos = 1; 2741 if (noneg && nopos) 2742 break; 2743 continue; 2744 } 2745 2746 /* 2747 * Follow the VM object chain to obtain the page to be mapped 2748 * into the pmap. This version of the prefault code only 2749 * works with terminal objects. 2750 * 2751 * WARNING! We cannot call swap_pager_unswapped() with a 2752 * shared token. 2753 */ 2754 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2755 2756 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2757 if (m == NULL || error) 2758 continue; 2759 2760 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2761 (m->flags & PG_FICTITIOUS) == 0 && 2762 ((m->flags & PG_SWAPPED) == 0 || 2763 (prot & VM_PROT_WRITE) == 0 || 2764 (fault_flags & VM_FAULT_DIRTY) == 0)) { 2765 /* 2766 * A fully valid page not undergoing soft I/O can 2767 * be immediately entered into the pmap. 2768 */ 2769 if ((m->queue - m->pc) == PQ_CACHE) 2770 vm_page_deactivate(m); 2771 if (prot & VM_PROT_WRITE) { 2772 vm_object_set_writeable_dirty(m->object); 2773 vm_set_nosync(m, entry); 2774 if (fault_flags & VM_FAULT_DIRTY) { 2775 vm_page_dirty(m); 2776 /*XXX*/ 2777 swap_pager_unswapped(m); 2778 } 2779 } 2780 pmap_enter(pmap, addr, m, prot, 0, entry); 2781 mycpu->gd_cnt.v_vm_faults++; 2782 if (curthread->td_lwp) 2783 ++curthread->td_lwp->lwp_ru.ru_minflt; 2784 } 2785 vm_page_wakeup(m); 2786 } 2787 } 2788